Re: [PATCH v2 net-next] net: socionext: get rid of huge dma sync in netsec_alloc_rx_data

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Fri, 10 Jan 2020 19:19:40 +0100
Lorenzo Bianconi <lorenzo.bianconi@redhat.com> wrote:

> > On Fri, 10 Jan 2020 16:34:13 +0100
> > Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> >   
> > > > On Fri, Jan 10, 2020 at 02:57:44PM +0100, Lorenzo Bianconi wrote:    
> > > > > Socionext driver can run on dma coherent and non-coherent devices.
> > > > > Get rid of huge dma_sync_single_for_device in netsec_alloc_rx_data since
> > > > > now the driver can let page_pool API to managed needed DMA sync
> > > > > 
> > > > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > > > > ---
> > > > > Changes since v1:
> > > > > - rely on original frame size for dma sync
> > > > > ---
> > > > >  drivers/net/ethernet/socionext/netsec.c | 43 +++++++++++++++----------
> > > > >  1 file changed, 26 insertions(+), 17 deletions(-)
> > > > >     
> > > 
> > > [...]
> > >   
> > > > > @@ -883,6 +881,8 @@ static u32 netsec_xdp_xmit_back(struct netsec_priv *priv, struct xdp_buff *xdp)
> > > > >  static u32 netsec_run_xdp(struct netsec_priv *priv, struct bpf_prog *prog,
> > > > >  			  struct xdp_buff *xdp)
> > > > >  {
> > > > > +	struct netsec_desc_ring *dring = &priv->desc_ring[NETSEC_RING_RX];
> > > > > +	unsigned int len = xdp->data_end - xdp->data;    
> > > > 
> > > > We need to account for XDP expanding the headers as well here. 
> > > > So something like max(xdp->data_end(before bpf), xdp->data_end(after bpf)) -
> > > > xdp->data (original)    
> > > 
> > > correct, the corner case that is not covered at the moment is when data_end is
> > > moved forward by the bpf program. I will fix it in v3. Thx  
> > 
> > Maybe we can simplify do:
> > 
> >  void *data_start = NETSEC_RXBUF_HEADROOM + xdp->data_hard_start;
> >  unsigned int len = xdp->data_end - data_start;
> >   
> 
> Hi Jesper,
> 
> please correct me if I am wrong but this seems to me the same as v2.

No, this is v2, where you do:
   len = xdp->data_end - xdp->data;

Maybe you mean v1? where you calc len like:
   len = xdp->data_end - xdp->data_hard_start;
   

> The leftover corner case is if xdp->data_end is moved 'forward' by
> the bpf program (I guess it is possible, right?). In this case we
> will not sync xdp->data_end(new) - xdp->data_end(old)

Currently xdp->data_end can only shrink (but I plan to extend it). Yes,
this corner case is left, but I don't think we need to handle it.  When
a BPF prog shrink xdp->data_end, then i believe it cannot change that
part the shunk part any longer.


> 
> > The cache-lines that need to be flushed/synced for_device is the area
> > used by NIC DMA engine.  We know it will always start at a certain
> > point (given driver configured hardware to this).


-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer